Psychological Methods

2002, Vol. 7, No. 1, 6482

Copyright 2002 by the American Psychological Association, Inc.

1082-989X/02/$5.00 DOI: 10.1037//1082-989X.7.1.64

Principles and Practice in Reporting Structural

Equation Analyses
Roderick P. McDonald and Moon-Ho Ringo Ho
University of Illinois at UrbanaChampaign
Principles for reporting analyses using structural equation modeling are reviewed,
with the goal of supplying readers with complete and accurate information. It is
recommended that every report give a detailed justification of the model used,
along with plausible alternatives and an account of identifiability. Nonnormality
and missing data problems should also be addressed. A complete set of parameters
and their standard errors is desirable, and it will often be convenient to supply the
correlation matrix and discrepancies, as well as goodness-of-fit indices, so that
readers can exercise independent critical judgment. A survey of fairly representative studies compares recent practice with the principles of reporting recommended
here.

those listed supply essentially the same basic information, with minor variations in the details supplied.
Thus, the eight parameter LISREL model, which
arose out of the work of Keesling and Wiley (see
Wiley, 1973) and was subsequently developed to
its current state by Joreskog (see Joreskog and Sorbom, 1996), the four-matrix model of Lohmoller
(1981), the three-matrix EQS model of Bentler and
Weeks (1980), and the two-matrix RAM model (see
McArdle & McDonald, 1984) rearrange the same set
of parameters. Not surprisinglyand perhaps not regrettablyuser guides and texts on this topic are not
in agreement in their recommendations about the
style of presentation of results (e.g., see Bollen, 1989;
Loehlin, 1992; Long, 1983a, 1983b). There is even
less agreement in the form of the results actually reported in articles on applications.
It would be immodest for any journal article to
offer a code of practice for the presentation of SEM
results. It could also be counterproductive. (We note
that for a long time there was a uniformly accepted
convention for the publication of analysis of variance,
or ANOVA results: the standard ANOVA table and
the table of cell means. The near-disappearance of this
from journals is regrettable.) Sound guidelines for the
reporting of SEM results have been offered previously
by Steiger (1988), Breckler (1990), Raykov, Tomer,
and Nesselroade (1991), Hoyle and Panter (1995), and
Boomsma (2000). MacCallum and Austin (2000) provided an excellent general survey of problems in applications of SEM.

Structural equation modeling (SEM), also known

as path analysis with latent variables, is now a regularly used method for representing dependency (arguably causal) relations in multivariate data in the
behavioral and social sciences. Following the seminal
work of Joreskog (1973), a number of models for
linear structural relations have been developed
(Bentler & Weeks, 1980; Lohmoller, 1981; McDonald, 1978), and work continues on these. Commercial
statistical packages include LISREL (Joreskog & Sorbom, 1989, 1996), EQS (Bentler, 1985, 1995), CALIS
(Hartmann, 1992), MPLUS (Muthen & Muthen,
1998), RAMONA (Browne, Mels, & Cowan, 1994),
SEPATH (Steiger, 1995), and AMOS (Arbuckle,
1997). Available freeware includes COSAN (Fraser
& McDonald, 1988) and Mx (Neale, 1997).
McArdle and McDonald (1984) proved that different matrix formulations of a path model with latent
variables are essentially equivalent. Programs such as
Roderick P. McDonald and Moon-Ho Ringo Ho, Department of Psychology, University of Illinois at Urbana
Champaign.
We thank Jack McArdle, Malgorzata Szewczyk, and Bob
Henson for their comments on this article. Any errors are
our responsibility.
Correspondence concerning this article should be addressed to Roderick P. McDonald, Department of Psychology, University of Illinois at UrbanaChampaign, 603
East Daniel Street, Champaign, Illinois 61820. E-mail:
rmcdonal@s.psych.uiuc.edu

64

REPORTING STRUCTURAL EQUATION ANALYSES

Our objective is to review some principles for reporting SEM results. We also summarize some observations on the variety of treatments of results given in
a selected set of reports on applications. The comparison of principles with practice is intended to underscore the importance of the recommendations made
here. These are aimed at increasing the information
supplied to the reader. Increased information should
allow a more critical assessment of the study reported
and serve to advance scientific knowledge. Naturally,
some of our recommendations are essentially endorsements of those already given. Furthermore, this topic
may need occasional revisiting until practice conforms with principles. We concentrate (appropriately)
on matters in which there should be little disagreement. However, we do address a number of fresh
elements here. We try to give fairly specific guidance
on implementing previously published recommendations as well as our own. We also offer a few mild
departures from previous suggestions. It is assumed
that the reader is familiar with basic SEM method and
with common terminology. However, we do offer occasional reminders about method, as well as remarks
about recent developments that the reader may have
missed.
From well-known SEM principles, we can formulate a list of results that we might hope to find in a
comprehensive report, and we can check current practice against this list. Of course, there can be a conflict
between an ideal of completeness and the understandable desire for conciseness in journal publications. This conflict is severe in the case of very large
models.
We surveyed articles using SEM from 1995 to 1997
in the following journals: British Journal of Psychology, Child Development, Developmental Psychology,
Journal of Abnormal Psychology, Journal of Applied
Psychology, Journal of Applied Social Psychology,
Journal of Consulting and Clinical Psychology, Journal of Counseling Psychology, Journal of Educational
Psychology, Journal of Family Psychology, Journal
of Personality and Social Psychology, Journal of Research in Personality, and Psychological Assessment.
The method of search was simply to look for all path
diagrams in the journals and period named. The intention was to find a reasonably representative
sample, limited to path models with latent variables
and to single-population studies.1 This survey yielded
100 possible articles for review, of which 41 met the
criteria: (a) path diagrams must be provided; (b) models being fitted must include both measurement and

65

structural parts (pure factor analytic models or path

analytic models without latent variables were excluded); and (c) multiple group comparisons were excluded (5 on this criterion).
The following discussion treats principles and practice together under familiar headings: model specification, identifiability, data and estimation, goodness
of fit, parameters and their standard errors and, finally, alternative models. Each section concludes with
specific recommendations.

Model Specification
Generally, a structural equation model is a complex composite statistical hypothesis. It consists of
two main parts: The measurement model represents a
set of p observable variables as multiple indicators
of a smaller set of m latent variables, which are usually common factors. The path model describes relations of dependencyusually accepted to be in some
sense causalbetween the latent variables. We reserve the term structural model here for the composite
SEM, the combined measurement and path models.
This avoids the ambiguity that arises when the path
model component is also labeled the structural
model.
In most applications the measurement model is a
conventional confirmatory factor model; the latent
variables are just common factors and the error or
specific terms are uncorrelated. The most common
exception concerns longitudinal models with measurements replicated at two or more time points. Longitudinal models usually need specific components
that are correlated across replicated measurements.
Very commonly, the measurement model is an independent clusters model, that is, a factor model in
which no indicator loads on more than one common
factor.
In turn, the path model structures the correlation or

Path models containing only observed variables

possibly weighted composite variables as in Wolds Partial
Least Squares (PLS)raise a separate set of issues from
those conventionally recognized in SEM (see McDonald,
1996, for an account of some of these). Multiple population
studies, likewise, raise questions of scale and of the structure of mean vectors, hence of goodness of fit, that would
make them difficult to integrate into this account. It should
be of interest to carry out reviews of articles written over
later periods in the hope of finding improvement.

66

MCDONALD AND HO

covariance matrix of the common factors. This structuring usually corresponds to a set of conjectured
causal relations. The path model itself is also a composite hypothesis. It requires the specification both of
a set of present versus absent directed arcs (paths)
between latent variables, and a set of present versus
absent nondirected arcs. Terminology in SEM is currently in a state of flux, because of the introduction of
ideas from directed acyclic graph theory (for a general
account, see Pearl, 1998, 2000). Here we follow Pearl
in referring to a single direct connection between two
variables in a path diagram as an arc, not a path. We
reserve the term path for a sequence of arcs connecting two variables.
Pearl (2000) made a strong case for the position
that a path model represents the operations of causality. The directed arcs indicate direct effects of conjectural (counterfactual) actions, interventions, or
states of the world, whereas nondirected arcs represent correlated disturbances, random terms corresponding to variations not explained by the model. A
directed arc in the graph of a path model is a singleheaded arrow from one variable to another. A nondirected arc is usually drawn as a double-headed arrow.
(The reader might wish to refer to Figures 13, where
directed arcs drawn from the latent variable F1 to F2
and to F3 and from F1 to the indicator variables Y1 and
Y2 are examples of directed arcs, whereas the doubleheaded arrow between F2 and F3 in Figure 1 or between d2 and d3 is an example of a nondirected arc in
Figure 2.) On the causal interpretation, the absence of
a directed arc from one variable to another implies the
absence of a direct effect, whereas the absence of a
nondirected arc implies that there are no omitted variables explaining their relationship. (These issues will
be explored in greater detail later.)
In contrast to path equations, regression equations
are essentially predictive and correspond to conditioning on observations of explanatory variables without
manipulationactual or theoretical. This is a rather
technical distinction. For our purposes we note that
residuals in a linear regression equation are uncorrelated with the independent variables by definition.
The disturbances (unexplained variations) in a path
equation can be correlated with the causal variables in
that equation. Pearl (2000) gave graph-theory conditions under which a path equation is a regression
equation; McDonald (in press) gives algebraic conditions. Most applications use recursive models, with no
closed cycles formed by directed paths. With commonly made assumptions, the path equations of re-

cursive models satisfy the conditions for regression

equations. In nonrecursive models there are directed
paths consisting of a sequence of unidirectional arcs
forming closed loops. The most common form of loop
is the simple case in which there is a directed arc from
variable X to variable Y and also from Y to X. More
rarely, we may find cases in which there is a directed
arc from X to Y, Y to Z, and Z to X, closing the loop.
In applications, loops formed by more than three variables seem rare to the point of nonexistence. Generally, their path equations are not regression equations.
For a nontechnical account of this point, see McDonald (1997).
On this causal view of path models, the specification of directed arcs rests on a substantive theory expressed as a rich set of causal (counterfactual) conjectures. Investigators of an empiricist persuasion
may hold alternative theoriesexplicit or implicit
justifying the direction of a directed arc, without acknowledging a notion of causal effect. We cannot
legislate this question in the current state of philosophical debate.
Ideally, we might hope that a published report gives
theoretical (causal?) grounds for the presence or absence of every directed arc in the path model. This
was strongly suggested by Hoyle and Panter (1995)
and Boomsma (2000), and we endorse it. A careful
reading of the selected studies suggests that the researchers chose to omit or include directed arcs with
some substantive justification. We cannot question
these choices, but we note that none of the 41 reports
examined attempted to approach the ideal of accounting for every directed arc that was chosen or omitted.
In the absence of an agreed methodology, some compromise seems inevitable in practice. And in the absence of the authors explicit account of the theoretical justification of the choice of each directed arc, the
reader is usually responsible for checking the plausibility of the set of choices made. We are not in a
position to apply the more tolerant criterion of plausibility across the wide range of applications examined. In the absence of a detailed account of the
choices made, it is tempting to surmise that these
could appear somewhat arbitrary, even to experts in
the relevant theory.
Quite generally in applications, the choice of nondirected arcs lacks any explicit justification in the
published reports. On reflection, it is easy to see that
the decision to omit a nondirected arc requires as
much theoretical justification as the decision to omit a
directed arc. The omission of a directed arc corre-

REPORTING STRUCTURAL EQUATION ANALYSES

sponds to a belief that there is no direct causeeffect

relation between two variables. The omission of a
nondirected arc corresponds to the belief that the variables have no common cause that has been omitted
from the model (e.g., see McDonald, 1997, 1999,
chap. 17; Pearl, 1998, 2000). Indeed, the possibility of
unspecified omitted common causes is the Achilles
heel of SEM. The substantive motivation for including a nondirected arc seems to rest on the belief that
some unmeasured common cause creates an unexplained relationship between them.
We note that two opposite research strategies suggest themselves: (a) We can omit a nondirected arc,
unless we are confident that an omitted common
cause exists; and (b) we can include the arc unless we
are confident that no such cause exists. The first strategy seems almost universally adopted in published
applications, but the choice is not made explicit.
Omission of nondirected arcs combines with the
omission of directed arcs to make a testable model.
With appropriate specification, the combination also
makes an identified model. The first strategy, in the
extreme, is a likely source of poorly fitting models.
The second strategy, in the extreme, yields untestable,
underidentified models. Even so, testability and identifiability should not be the primary motives for omitting parameters. To underline the dilemma facing the
investigator we note that adding nondirected arcs does
not change the causal hypothesis, yet it radically alters
fit.
Short of the ideal of a complete accounting, we
simply suggest what should be obvious, namely, that
the theoretical justification of directed arcs should be
given as much detail as possible, hence open to the
plausibility judgment of a reader with relevant substantive expertise. The justification for omission or
inclusion of nondirected arcs (correlated disturbances)
seems a more difficult research task. Omission might
be covered by a blanket admission of ignorance. This
carries with it the recognition that all conclusions
could be radically altered by adding relevant variables
in future studies. Inclusion requires specific theoretical grounds. We will return to this issue in other contexts.

Identifiability
Desirably, the SEM report will contain an explicit
account of the conditions on the model that will secure identifiability, which is logically prior to estimation. Pearl (1998, 2000) gave graphical conditions for

67

the identifiability of directed arc or path coefficients

in the path model, and McDonald (in press) gives
corresponding algebraic conditions. (We say that the
model is identified if every parameter is identified.)
Bekker, Merckens, and Wansbeek (1994) gave very
general, very technical methods for determining
model identification. However, in most applications
these technicalities can be avoided. Commonly, we
just require a fairly simple treatment of three distinct
problems, namely, identifiability of the measurement
model, identifiability of the path model, and scaling of
the latent variables. We now examine these in turn.

Identifiability of the Measurement Model

A known, strong, sufficient condition for the parameters of the measurement model to be identified
(except for scaling) is the condition that the factor
loadings form independent clusters. In an independent
clusters model each observed variable loads on only
one common factor, and we can call it a pure indicator of the factor. A weaker sufficient condition has
been called an independent clusters basis by McDonald (1999). This requires each latent variable or common factor to have at least two pure indicators if the
factors are correlated, and at least three if they are not.
Fortunately, in applications, this condition (and the
stronger independent clusters condition) commonly
results from substantive considerations. The substantive motivation for choosing the pattern of zero versus
nonzero factor loadings dates from the origins of the
common factor model. By design of the measurements we can usually expect that a subset of indicator
variables measures just one attribute of the examinees
in common, although each also measures a unique
component (and is possibly subject to an error of replication). Such a design will tend to give independent
clusters. It should be noted that the classical Thurstonian conditions of simple structure are often confused with independent clusters. Simple structure is
more general. It allows the possibility that every variable loads on more than one factor. Simple structure
does not generally secure an identified measurement
model, although many simple structures will do so.
Of the 41 studies in our survey, 25 have independent cluster structure, 2 have an independent clusters
basis, and 8 contain mixed latent and observable variables in the path model, but are readily recognized to
be identified by independent clusters criteria. Only 3
studies do not have a measurement model that is seen
to be identified by inspection. Nevertheless, in nearly
every case, the reader is responsible for verifying

68

MCDONALD AND HO

identifiability because no account is provided in the

report.
The recommendation that follows from these considerations is this: Authors have a duty to show evidence that the measurement model is identified. Commonly, this will be the presence of independent
clusters, or at least an independent clusters basis. If
authors fail to do this, the reader can usually check the
point without inconvenience.

Identifiability of the Path Model

The identifiability of the path model rests crucially
on the choice of nondirected arcs. To examine this
problem, we first recall the distinction between exogenous and endogenous variables, and introduce a further useful distinction. In terms of the path diagram,
an exogenous variable has no directed arc ending on
it. Its entire variance is unexplained by variables in
the set studied. An endogenous variable has at least
one directed arc ending on it, originating from one or
more exogenous or endogenous variables. (Again, see
Figures 13 for illustrations; F1 is an exogenous latent
variable, whereas F2 and F3 are endogenous.) There is
a further useful distinction concerning the order of the
variables in the path model. A variable X precedes
another variable Y if there is a directed path (a sequence of unidirectional arcs) from X to Y. Two such
variables are causally ordered. The variables are fully
ordered if we can list them so that there is a directed
path to each variable from every variable listed before
it. (In Figures 13, F2 and F3 are unordered, whereas
F1 precedes both F2 and F3, so F1 and F2, also F1 and
F3 are causally ordered.)
A weak condition for the identifiability of a recursive path model is that every equation is a regression.
This condition is difficult to test directly (see McDonald, 1997, in press). Fortunately, it is implied by the
stronger condition that all covariances of disturbances
of causally ordered variables are zero (McDonald,
1997, termed this the precedence rule). This condition
is easily checked from the path diagram. The precedence rule is implied by the stronger condition that all
disturbances of endogenous variables are uncorrelated
(McDonald, 1997, termed this the orthogonality rule).
If a model is fully ordered except for the exogenous
variables, the precedence and orthogonality rules coincide. These rules are not well and widely known.
There is a commonly accepted view that in a recursive path model identifiability may or should be
secured by the orthogonality rule, but in a nonrecursive model the orthogonality rule cannot be em-

ployed. Instead, a rather technical inquiry is suggested. With all disturbances allowed to be nonzero,
the rank and order rules are applied to see if the exogenous variables can yield identifiability by serving
as instruments. These rules are too technical to be
described here; for an introduction to this method, see
Bollen (1989). (We note that if we follow these rules,
the covariances between disturbances cannot be accounted for by omitted variables.) The contrary view
is that we may assume orthogonality in nonrecursive
models. This belief is at least implicit in the work of
a number of authors. For example, Spirtes, Richardson, Meek, Scheines, and Glymour (1998) used it in a
graph-theory account, and MacCallum, Wegener,
Uchino, and Fabrigar (1993) used it in an account of
equivalent models.2 There appear to be unresolved
problems in the foundations of nonrecursive models.
These have been interpreted in a number of ways,
including feedback, systems measured in equilibrium, or averaged time series (see McDonald, 1997,
for a nontechnical account of some of these issues,
and Fisher, 1970, for a technical account of the averaged time series; see, also, Gollob & Reichardt, 1987,
for the problem of time lags in such models). In the
current state of knowledge we cannot adjudicate between these views, but it is fair to state that it is much
easier to draw a nonrecursive loop in a path model
than to motivate it rigorously.
Of the 41 studies surveyed, 10 have fully ordered
models (except exogenous variables). Of these, all
except 1 followed the orthogonality rule (and, equivalently, the precedence rule). The exception gave no
substantive reason for including a nondirected arc between two variables connected by a directed arc (and
no discussion of a likely identifiability problem). Of
the remaining 31 studies that have one or more unordered pairs of latent variables, 7 chose nondirected
arcs for the unordered pairs, as if they were following
the precedence rule, and 24 had no correlated distur-

These opposing views seem to have arisen because the

assumption that every path equation is a regression is implied by the assumption that the disturbances are orthogonal
if the model is recursive, and contradicted by the latter
assumption if the model is nonrecursive. The contradiction
can be resolved by denying either assumption. The instrumental variables treatment abandons both assumptions in
favor of the assumption that the exogenous variables are
orthogonal to the disturbances. It does not seem that any
account of foundations has been given to adjudicate between these choices (see McDonald, 1997).

REPORTING STRUCTURAL EQUATION ANALYSES

bances, as if they were following the orthogonality

rule. Six longitudinal studies appropriately included
correlated disturbances corresponding to replicated
observations. As already noted, explicit reasons for
the choices to include or to omit nondirected arcs are
lacking in all of the studies.
Our recommendation on this matter is as follows:
In the common case of a recursive model, the choice
to include or omit a nondirected arc should, in principle, rest on substantive grounds for supposing the
existence or nonexistence of unmeasured common
causes, and not on securing identifiability. If the resulting model has no nondirected arcs between causally ordered variables, it is certainly identified. (This
principle is not widely recognized and may not be
known to the general user.) In a case with a nondirected arc between ordered variables, investigation is
desirable using the rather technical methods of Pearl
(1998, 2000), McDonald (in press), or Bekker et al.
(1994). In the case of a nonrecursive model there is
currently no clear identifiability rule. Perhaps there is
also no single clear motive for the use of such a
model.

Identifiability and Scaling

It is easy to choose the scale of exogenous latent
variables: simply set their variances to unity in the
computer program. The variance of an endogenous
variable is not a parameter in the model, so fixing the
scale on which it is measured becomes problematic.
Currently, there are three main procedures for dealing
with this problem. Browne and Du Toits (1992)
general method can be used to set all variances of
endogenous latent variables to unity by a constrained
minimization procedure. McDonald, Parker, and Ishizukas (1993) reparameterization method also sets
these variances to unity, but is restricted to recursive
models. It is also possible to scale by setting chosen
factor loadings or error variances to a constant (usually unity) during minimization and then to rescale the
latent variables after fitting it. Browne and Du Toits
method is implemented in RAMONA and SEPATH;
McDonald et al.s method can be implemented in
PROC CALIS, COSAN, or any program allowing
suitable FORTRAN-like supplementary programming. The rescaling method is implemented in
LISREL and EQS. The first two methods share the
advantage that they yield correct standard errors in a
fully standardized model. The rescaling method obtains correct parameters in a completely standardized
model, but without corresponding standard errors.

69

Of the 41 reports in the survey, 40 gave standardized solutions, the exception being also the only nonrecursive model. All these are presumably computed
by the rescaling method, as they all used either
LISREL or EQS for the analyses.
It is certainly an old psychometric custom to interpret the numerical values of standardized factor loadings, standardized regression coefficients, and standardized path coefficients. It seems as though these
are often thought to be metrically comparable, although their unstandardized counterparts are clearly
not. This is an unsuitable place for a lengthy discussion of the dilemmas underlying the use of standardized coefficients, unstandardized coefficients, or variance explained as measures of the importance of an
explanatory variable in a regression or in a path
model.
We accept that standardization either before or after
estimation is virtually unavoidable for applications of
a path model with latent variables. Standardization
avoids underidentifiability due to arbitrariness of
scale. Experience seems to show that a completely
standardized solution also aids interpretation of the
results. As noted, some computer software is available
that uses methods for obtaining standardized solutions
with correct standard errors. We hope that the use of
these methods will increase in the near future.

Data and Estimation

In the commonly employed estimation procedures,
a sample of observations of size N on p variables
gives a p p sample covariance or correlation matrix
S, with elements sjk. A computer program then minimizes some function of the discrepancies sjk jk
between these and the fitted covariances jk implied by the composite structural model. Possibilities
include maximum likelihood (ML), ordinary (unweighted) least squares (OLS), generalized least
squares (GLS), and a variety of weighted least squares
(WLS) discrepancy functions intended to give good
estimates without requiring multivariate normality;
these last stem from the seminal work of Browne
(1984).

Multivariate Normality
Both ML and GLS estimation in SEM require the
assumption of multivariate normality. However, as
Micceri (1989) suggested, much social and behavioral
science data may fail to satisfy this assumption. Several studies of the robustness of the multivariate nor-

70

MCDONALD AND HO

mality assumption (Amemiya & T. W. Anderson,

1990; T. W. Anderson, 1989; Browne & Shapiro,
1988; Satorra & Bentler, 1994) have found that parameter estimates remain valid under reasonable assumptions even when the data are nonnormal,
whereas standard errors do not. (These assumptions
are difficult to test in applications.) A number of
simulation studies (Chou, Bentler, & Satorra, 1991;
Hu & Bentler, 1995; West, Finch, & Curran, 1995)
suggest that ML and GLS estimation can give biased
standard errors and incorrect test statistics in the presence of excessive skewness and/or kurtosis in the
data.
The multivariate normality assumption may be
evaluated univariately by checking the marginal distribution of each variable or by Mardias (1970) multivariate skewness and kurtosis coefficients. Outliers,
a likely source of skewed data, can be detected from
the univariate distribution of the variables. Bollen and
Arminger (1991) suggested the use of factor scores
for outlier detection. Of the 41 studies, just 5 reported
a test of multivariate normality in justification of the
ML procedure, applied in all cases.
To deal with nonnormal data, Browne (1984) developed an asymptotically distribution-free (ADF) estimator, but a very large sample size is needed to
obtain reliable weight matrices. For elliptical distributions, Browne (1984), Kano (1992), Shapiro and
Browne (1987), and Satorra and Bentler (1994) have
developed corrections to the normal likelihood ratio
statistic. (To avoid technicalities, we simply remark
that if a set of variables follows an elliptical distribution, all its marginal distributions are symmetric with
the same kurtosis. Estimation based on an elliptical
distribution can be thought of as intermediate between
estimation using normality and ADF estimation.)
Simulation studies by Curran, West, and Finch
(1996), and by Hu, Bentler, and Kano (1992) suggest
that the Satorra and Bentler rescaled statistic works
well over a variety of distributions (see, also, Bentler
& Dudgeon, 1996). In the context of analysis of variance and multivariate regression problems, transformation of variables is a common method for dealing
with nonnormality of data. Mooijaart (1993) proposed
the use of univariate Box-Cox transformations. Yuan,
Chan, and Bentler (2000) proposed a robust transformation method.
The presence of categorical variables or indicators
may cause nonnormality. Muthen (1984) developed a
continuous/categorical variable methodology (CVM)
estimator, which allows the analysis of any combina-

tion of dichotomous, ordered polytomous, and measured variables. Like Brownes (1984) ADF estimator, it requires a very large sample size to obtain
reliable weight matrices. Simulation studies (Muthen,
1989; Muthen & Kaplan, 1992) suggest that the CVM
estimator outperforms the Satorra and Bentler (1994)
and ADF estimators when the number of categories of
the variables are few (< 5).
A mild dilemma stems from the fact that ML estimation and its associated statistics seem fairly robust
against violations of normality, whereas the use of
ADF estimators requires extremely large samples for
reliable weight matrices, far larger than are commonly
available in current SEM applications.
Accordingly, we hesitate to make firm recommendations for the resolution of this dilemma, beyond
noting that Mardias (1970) test of multivariate skewness and kurtosis is well known and implemented in
available computer software. It should, therefore, be
easy for the investigator to see if a problem appears to
exist and to report it to the reader. But in many cases
the sample size will require the investigator to rely on
the robustness of ML/GLS methods.

Sampling and Missing Data

A common problem in SEM applications is missing
data. There are many possible reasons why the data
matrix may be incomplete. Rubin (1976) proposed a
classification scheme for missing data mechanisms
and argued that missing data can be ignored (i.e.,
unbiased estimates can be obtained) under the condition that data are missing completely at random
(MCAR) and missing at random (MAR). MCAR refers to missing data on a variable where presence or
absence of the observation is independent of other
observed variables and the variable itself. This is a
stringent assumption that may not be justifiable in
practice (Muthen, Kaplan, & Hollis, 1987). Less restrictively, MAR allows the presence or absence of an
observation to depend on other observable variables,
but not the variable from which the value is missing.
Listwise and pairwise deletion for missing data can
lead to biased parameter estimates under MAR, but
unbiased parameter estimates can be obtained under
MCAR. Under MCAR and MAR, unbiased and more
efficient parameter estimates can be obtained by full
information maximum likelihood estimation (FIML),
also known as the individual raw-score likelihood
method (for simulation studies, see, e.g., Enders &
Bandalos, 2001). Alternatively, if there are not too
many distinct patterns of missing data, multiple-group

REPORTING STRUCTURAL EQUATION ANALYSES

SEM may be used. This requires a different group for

each pattern of missing data, with equality constraints
imposed across groups. The implementation of this
method can be found in Allison (1987), McArdle
(1994), and Muthen et al. (1987). The expectation
maximization (EM) algorithm (Dempster, Laird, &
Rubin, 1977) may also be employed. Recently, multiple imputation has become a popular method to deal
with missing data. Multiple data sets are created with
a set of plausible values replacing the missing values,
then standard analysis is applied to the complete data
set (for details of this procedure, see Graham & Hofer,
2000). Currently, available software for dealing with
missing data includes Mplus, AMOS, and Mx, using
full information, whereas LISREL provides full information and multiple imputation; all of these assume
MAR.
If the missing data mechanism is nonignorable (not
MAR), there is no correction available. We may guess
that this will be the common case in applications (for
further discussion, see Heckman, 1979; Heckman &
Robb, 1986; McArdle, 1994; Rubin, 1987, 1991).
The problem of missing data is closely related to
the problem of sampling as such in SEM studies. In
principle, we scale a measurement model by choosing
a zero mean and unit standard deviation for the latent
variables in the population sampled. The likely effect
of nonrandom sampling (which is surely common in
applications) is to change (in an unknown way) the
variances and covariances of the latent variables. The
metric of the model will then fail to correspond to any
well-defined norming population. However, in many
applications there is no defined norming population
in the first place. For applications such as testing a
causal model, this may not be critical. Random sampling will be of primary importance mainly in certain
test theory applications, where tests are scaled on the
basis of a reference or calibration population. In such
cases there will likely be no accompanying causal
model, the object being primary test development (see
McDonald, 1999).
The effect of nonrandom sampling includes the effect of listwise deletion. Technically, the likely effect
of either is to produce a change (of unknown extent)
in the variances and covariances of the latent variables. Because we scale the latent variables to have
unit variance, the result will be proportional changes
(again of unknown extent) in factor loadings. This
expectation follows from the classical account by
Meredith (1964) of the effect of selection on a factor
model. We hesitate to make firm recommendations on

71

the problem of missing data, beyond the obvious point

that a report should indicate the extent to which there
is a missing data problem and describe the method
used to deal with it. The report should also include
some account of the extent to which nonrandom sampling (whether caused by missing data or not) could
significantly work against the purposes of the study.

Reporting Data
Subject to editors concerns about space, the
sample covariance or correlation matrix gives the
reader a great deal of freedom to formulate and evaluate plausible alternative models. Mere inspection of
this information can often be useful to the reader.
Either the sample correlation matrix with or without standard deviations or the sample covariance matrix was supplied in 19 of the 41 reports (M 14.74
variables, SD 6.22, range 928). The 22 that did
not supply covariances or correlations have a mean of
21.74 (SD 8.55, range 1040), with only 4 in
excess of 28; none of the latter indicated availability
of their data.
It is desirable that this information be readily available. Few authors justify their model in detail, and it
is generally easy to postulate equally plausible, and
possibly better-fitting alternatives on the basis of such
theory as is available. (Any instructor in quantitative
methods who has given this task as an exercise to
graduate students will confirm this statement.) We
suggest that there is a strong case for publishing the
correlations, possibly accompanied by means and
standard deviations, for up to 30 variables. In the case
of a large set of variables, the author can alternatively
indicate a means to access the covariance or correlation matrix if the reader wishes to do so; possibilities
include application to the author or placing the information in the authors, or possibly the journals, website. In the case of truly large studies, a more comprehensive account than is possible in a journal might
be made available in this way. Information technology is changing too rapidly for this advice to be more
precise.

Goodness of Fit
Except for OLS estimates as usually treated, estimation methods also yield an asymptotic chi-square
and asymptotic standard errors for the parameters in
an identified model. It has long been recognized that
all SEMs are simplified approximations to reality, not
hypotheses that might possibly be true. Accordingly,

72

MCDONALD AND HO

an abundance of indices has been developed as measures of the goodness or badness of the approximation
to the distribution from which the sample was drawn,
and it is very easy to invent many more (see, e.g.,
Bollen & Long, 1993; McDonald & Marsh, 1990).
Available computer programs supply varying subsets
of these. We may distinguish absolute and relative fit
indices: Absolute indices are functions of the discrepancies (and sample size and degrees of freedom);
relative indices compare a function of the discrepancies from the fitted model to a function of the discrepancies from a null model. The latter is almost
always the hypothesis that all variables are uncorrelated. Some psychometric theorists have prescribed
criterion levels of fit indices for a decision to regard
the approximation of the model to reality as in some
sense close. This is considered to make the decision
objective (see, e.g., Hu & Bentler, 1999). There are
four known problems with fit indices. First, there is no
established empirical or mathematical basis for their
use. Second, no compelling foundation has been offered for choosing a relative fit index over an absolute
index, or for regarding uncorrelated variables as a null
model. Third, there is not a sufficiently strong correspondence between alternative fit indices for a decision based on one to be consistent with a decision
based on another; the availability of so many could
license a choice of the best-looking in an application,
although we may hope this does not happen. Fourth,
and perhaps most important, a given degree of global
misfit can originate from a correctable misspecification giving a few large discrepancies, or it can be due
to a general scatter of discrepancies not associated
with any particular misspecification. Clear misspecifications can be masked by the indexed fit of the
composite structural model. It is impossible to determine which aspects of the composite hypothesis can
be considered acceptable from the fit indices alone.
Along with checking these, we recommend examining the (standardized) discrepancies in the measurement model and the individual discrepancies between
the latent variable correlations in the measurement
model and the fitted correlations constrained by the
path model.

Global Fit Indices

As we would expect, the question of goodness of fit
is resolved by different investigators in quite different
ways. Unsurprisingly, all 41 studies report the global
chi-square and degrees of freedom for the composite
structural model (the measurement model and path

model combined). Of these, 5 were in the enviable

position of having a conventionally nonsignificant
chi-square, at sample sizes 70, 165, 193, 330, and 461.
However, these, as well as the others, also reported
some indices of goodness or badness of global approximation. In terms of simple popularity, the index
independently given by McDonald and Marsh (1990)
as the unbiased relative fit index (URFI) and Bentler
(1990) as the comparative fit index (CFI)3 was most
used (21 of 41 studies) and next, used in 20 of 41
studies, was the root mean square error of approximation (RMSEA), which originated with Steiger and
Lind, and is accessible in published work by Browne
and Cudeck (1993). Other indices commonly used are
the goodness-of-fit index (GFI; Joreskog & Sorbom,
1989; 15 of 21 studies), the Tucker-Lewis index (TLI;
Tucker & Lewis, 1973; sometimes referred to as the
nonnormed fit index, or NNFI, attributed to Bentler &
Bonett, 1980; 13 of 41 studies), and the normed fit
index (NFI; Bentler & Bonett, 1980; 9 of 41 studies).
Most investigators reported at least two such measures. Those relying on the RMSEA generally accepted the authoritative claim that an RMSEA less
than .05 corresponds to a good fit and an RMSEA
less than .08 corresponds to an acceptable fit. Most
authors who used an index scaled up to unity for
perfect fit regarded these (URFI/CFI, GFI, etc.) as
acceptable if they were greater than .9. It is sometimes
suggested that we should report a large number of
these indices, apparently because we do not know
how to use any of them. It appears that the only metacriterion we can use to evaluate these conventions is
the principle that a model is acceptable if the discrepancies are too small to support a more complex model
(see McDonald, 1999, chap. 17). None of these studies give discrepancies (or, it seems, use their distributions to aid judgment), so we cannot check them
against such a criterion. Simulation studies based on
restrictive models do not help because the fit indices
were invented to deal with the fact that no restrictive
model fits real data.
If an article does supply the sample correlations, it

There is a minor difference between the CFI and the

URFI. If the URFI exceeds unity, corresponding to an overfitting model, the CFI is reset to unity, leaving the extent of
overfit unknowable. Hoyle and Panter (1995) gave this difference as a reason for preferring the CFI. We mildly demur, suggesting it is a reason for preferring the URFI as
being more informative.

REPORTING STRUCTURAL EQUATION ANALYSES

is very easy for the authors to complete the table by

including the standardized discrepancies in the
complementary triangle. This enables the reader to
form an independent judgment of the relationships not
well explained by the model. It becomes possible to
judge whether a marginal or low index of fit is due to
a correctable misspecification of the model, or to a
scatter of discrepancies, which suggests that the
model is possibly the best available approximation to
reality. If a covariance matrix has been analyzed, the
computer program may supply variance-standardized
discrepancies for this purpose.
Table 1 illustrates the format of a table presenting
correlations, variances, and standardized discrepancies, and is adapted from McDonald (1999, pp. 388
389), where details can be found. Such a table presents the information in a very compact form. We
note that in this (empirical) example the two nonzero
discrepancies correspond directly to missing arcs between the pairs of variables. The remaining discrepancies are necessarily exact zeros in this model, therefore, a global fit index would not seem very helpful.
For large numbers of variables the report can give
summary information about the discrepancies, from
which the reader can endorse or disagree with the
authors judgment. Listings of the largest discrepancies are given by some computer programs, and in any
case are easy for authors to create. Of the 19 of 41
studies giving the correlation matrix, none gave the
discrepancies or summary information about them.
(Given all the parameters, the reader can construct the
discrepancies, but not without considerable effort.)
Our comments on this issue might seem intended to
discourage the use of fit indices. Our intention is,
instead, to warn that the issue is unsettled. More constructively, but tentatively, we offer some recommendations as follows: It is our belief that no global index
of fit (together with a criterion for its acceptability)
Table 1
An Example of Reporting Correlations and Discrepancies
Correlations and discrepancies
5.86
.370
.400
.320
.210

.0
6.08
.290
.420
.410

.0
.0
3.00
.550
.630

.0
.0
.0
7.90
.730

.097
.084
.0
.0
7.54

Note. Sample correlations are in lower triangle; variances are in

diagonal and discrepancies are in upper triangle. An exact zero is
represented by .0. (Example adapted from McDonald, 1999, pp.
388389.)

73

can substitute for a detailed examination of the discrepancies. However, if inspection shows that these
are well scattered, they are adequately summarized in
the root mean square residual (RMR), which is an
immediately interpretable measure of the discrepancies. In turn, the GFI is a function of the RMR and the
corresponding root mean square of the sample correlations. The GFI is therefore acceptable to those who
believe that for a given RMR, fit is better if the correlations explained are larger. As shown in some detail by McDonald and Marsh (1990), most other fit
indices can be expressed as functions of the noncentrality parameter connected to ML/GLS estimation.
An unbiased estimate of the noncentrality parameter
is given by d (2 df )/N (McDonald, 1989). This
parameter is also a norm on the sizes of the discrepancies. Accordingly, if the discrepancies are well scattered, such indices capture their general spread well
enough. Those indices that have been shown (as in
McDonald & Marsh, 1990) to be free of sampling
bias, for example, the RMSEA and the URFI (and CFI
if not reset to unity because of overfitting), can be
recommended as supplements to the investigators
primary judgment based on the discrepancy matrix.

Path Model Fit

As already noted, a structural model is a composite
of a measurement model and a path (causal) model.
Accordingly, it might be useful to separate measures
of fit into parts corresponding at least to these two
major components. Surely the primary objective of an
SEM study is to give supporting evidence for the
specified path model. The (hopefully well designed)
set of measurements is essentially subservient to this
aim.
J. C. Anderson and Gerbing (1988) suggested a sequential testing procedure based on the recognition
that the structural model is nested within the measurement model. (See also Fornell & Yi, 1992, and the
reply by J. C. Anderson & Gerbing, 1992, as well as
Bentler, 2000; Bollen, 2000; Hayduk & Glaser,
2000a, 2000b; Herting & Costner, 2000; Mulaik &
Millsap, 2000, for further discussion of this question.)
The asymptotic distribution of the ML or GLS discrepancy function for a composite structural model
can be decomposed into independent additive noncentral chi-squares, one for the measurement model, and
one for the path model (see Steiger, Shapiro, &
Browne, 1985). The path model component is estimated as the difference between the discrepancy functions for the structural model and the measurement

74

MCDONALD AND HO

model. The degrees of freedom are correspondingly

additive, of course.
Separate chi-squares and degrees of freedom for the
measurement model, with unconstrained latent variable correlations, were given in 14 of the 41 studies.
These studies all provided some comparison of the

global and measurement chi-squares, but their final

conclusions were based only on the goodness of approximation of the composite structural model.
Table 2 gives the chi-squares and degrees of freedom for the structural and measurement model and
also the chi-squares and degrees of freedom for the

Table 2
Reexamination of the Goodness of Fit of the Structural (s), Measurement (m), and
Path (p) Models for 14 Selected Studies
N

Modela

df

RMSEA

461

357

465

326

507

81

237

289

330

10

197

11

377

12

1556

13

84

14

70

124.8
27.7
97.1
980.7
638.1
342.6
1368.2
1284.1
84.1
212.0
158.0
54.0
400.4
306.8
93.6
112.0
51.3
60.8
519.2
514.8
4.6
521.0
434.9
86.1
288.5
284.4
4.0
685.8
562.3
33.5
161.1
141.2
19.9
725.0
577.0
148.0
50.2
44.7
5.5
41.9
12.4
29.5

106
97
9
391
361
30
472
467
5
89
81
8
161
155
6
56
42
14
340
333
7
180
175
5
214
201
13
300
270
30
80
79
1
269
247
22
43
38
5
21
17
4

.001
1.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.000
.153
.000
.000
.000
.727
.000
.000
.000
.000
.000
.991
.000
.000
.299
.000
.000
.000
.000
.000
.000
.209
.211
.356
.000
.000
.000

.041
.150
.191
1.651
.678
.876
1.929
1.757
.170
.377
.236
.141
1.180
.299
.173
.691
.115
.576
.756
.767
.011
1.180
.899
.280
.226
.252
.026
1.958
1.940
.018
.215
.165
.050
.293
.212
.081
.086
.080
.006
.298
.066
.365

.020

s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p
s
m
p

Study

Note.

d noncentrality parameter; RMSEA root mean square error of approximation.

.146
.065
.046
.171
.064
.061
.184
.065
.054
.133
.054
.044
.170
.111
.052
.203
.047
.048
.081
.072
.237
.032
.035
.081
.085
.025
.052
.046
.224
.033
.029
.061
.045
.046
.035
.119
.302

REPORTING STRUCTURAL EQUATION ANALYSES

path model, obtained by difference, for the 14 studies

in our survey that provided the necessary information.
Not surprisingly, the degrees of freedom associated
with the path model are generally much smaller than
the degrees of freedom for the measurement model.
Accordingly, the fit of the composite structural model
can appear satisfactory when the few constraints implied by the path model are not, in fact, correctly
specified. Conversely, if the fit of the composite
model appears unacceptable, it is important to know if
the misfit is due primarily to a misspecified measurement model, a misspecified path model, or both.
As previously noted, the URFI (McDonald &
Marsh, 1990), the equivalent CFI (Bentler, 1990), and
most current fit indices are functions of the noncentrality parameter. This parameter also has an additive
property. The noncentrality of the structural model is
the sum of the noncentralities of the measurement and
the path models. Table 2 includes the unbiased estimates of the (additive) noncentralities as well as the
resulting RMSEAs (which, of course, are not additive) for these studies. If we are willing to rest on the
established properties of the RMSEA, the results suggest that in all but a few cases the goodness of fit of
the composite structural model, with its large number
of degrees of freedom, conceals the badness of fit of
the path model, with its much smaller number of degrees of freedom. By a natural extension of theory, we
may regard the additive noncentrality parameters
from the measurement and path model discrepancy
functions as representing the errors of approximation
of these two independent components of the model.
We tentatively conclude that in the majority of studies
for which the measurement model information is
available, the goodness of approximation of the path
model may be unacceptable, contrary to the published
conclusions. By extrapolation, this may be true of
those studies for which this information is not available.
It certainly seems desirable that existing SEM studies be reevaluated with attention to direct assessment
of the fit of the path model. For this purpose, we
might use a supplementary analysis by a two-stage
procedure. We first fit the measurement model, and
then fit the path model to the latent variable correlation matrix in order to study in detail the pattern of the
discrepancies. We recommend such a two-stage procedure in future studies.

Parameters and Standard Errors

The parameters of an SEM are the independently
estimated loadings and error variances and covari-

75

ances in the measurement model, and the independently estimated directed arc coefficients and disturbance variances and covariances in the path model.
Special cases include: (a) pure measurement models,
that is, traditional confirmatory common factor models; (b) path models for observable variables; and (c)
mixed models in which some variables in the path
model are observable and some are latent.
We note in passing that the possibility of mixed
models underlines a dilemma in SEM, needing further
research. As pointed out by McDonald (1996), the
researcher usually has a choice between making a
path model with latent variables and a path model
with composite variables: simple or weighted sums of
indicators. The common choice seems to be the
former, because it estimates causal relations between
attributes that are corrected for errors of measurement. For these causal relations to be taken seriously,
we must be able to suppose we could, in principle, add
enough further indicators to make error-free measurements. If the number of indicators of each attribute is
small (as seems common in applications), such corrections may themselves be unreliable. In a model in
which some attributes are measured by a single total
test score, we can take scores from items or subtests
of the test as multiple indicators, if we wish to allow
for errors of measurement. In the rare case where a
single item is the only measure of an attribute, there
does not yet seem to be any treatment available for its
error of measurement.
Just 12 of the 41 studies report all parameters,
whereas 20 omit both error and disturbance parameters, 2 omit error variances, 3 omit disturbance variances, 2 give only the directed arc coefficients, and 2
omit parameter estimates in the measurement part of
the model. Even in a fully standardized model, the
reader would find it difficult to construct the error
variances as unit complements of functions of the
other parameters. Generally, we would wish to verify
that the unique variances are not close to zero, corresponding to an improper (Heywood) solution, as
would also be indicated by large standard errors. The
disturbance variances also allow the reader to see
what proportions of variance of the endogenous variables are accounted for by the model (only 2 of the 41
reports explicitly addressed this question).
Standard errors of some parameters were reported
in only 5 of the 41 studies. Of these 5, none reported
all standard errors. Usually, standard errors of unique
and disturbance variances are not reported. Generally,
there appears to be nothing preventing such a report.

76

MCDONALD AND HO

Standard errors can be included in tables of the parameter values or, following one convention, put in
parentheses attached to the parameters in a path diagram. However, unless the scaling problem is solved,
either by the constrained minimization method of
Browne and Du Toit (1992) or by the reparameterization method of McDonald et al. (1993), standard errors are not available for the standardized solution. If
the standardized parameters are obtained by rescaling,
it would still be possible to examine and report the
standard errors of the unstandardized parameters, and
to describe statistical inferences based on these before
interpreting the corresponding standardized coefficients. This would follow the comparable procedure
in regression programs.
The obvious and conveniently implemented recommendation is to include all the parameters and their
standard errors in the research report. It is difficult to
see what could prevent this.
The parameters (and standard errors if supplied)
can, with equal convenience and approximately equal
space on the page, be presented in tables or attached
to directed and nondirected arcs in path diagrams.
Path diagrams originated with Sewell Wright (1921),
and have come into current SEM practice rather informally and haphazardly. There is at least some
agreement that the network of relationships in a path
model is most easily appreciated in the form of a path
diagram, a picture of the graph. Formal directed acyclic graph (DAG) theory (see Pearl, 1988, 2000;
Scheines, Spirtes, Glymour, Meek, & Richardson,
1998) is becoming more widely recognized by SEM
researchers. Recognition of this work should soon
lead to a greater formality in the pictorial representation of the hybrid graphs (graphs containing both directed and nondirected arcs) that structure SEMs, and
to their direct use in determining the precise form of
the constraints on the covariances.
We comment, and invite possible disagreement,
that in the common case where the measurement
model is a standard independent cluster model, there
is little to be gained from including the relations between common factors or latent variables and their
indicators in the path diagram. A good case can be
made for presenting the measurement model in the
tabular form of the older factor analysis tradition, and
drawing an uncluttered path diagram containing only
the latent variables (or any observable variables included in the path model), with their arcs and disturbance terms. This allows a much clearer appreciation
of the relations postulated in the path model than if the

diagram is complicated by the measurement relationships.

It cannot be claimed that there is a right way to
draw a path diagram. There appear to be three main
conventions currently available, in addition to the
simple picture of a graph favored by DAG theorists
(e.g., Pearl, 2000). (A number of minor variants in use
can be characterized by incompleteness.)
Figures 13, which are largely self-explanatory, illustrate these. (In applications, a diagram of the theoretical model also contains parameter names close to
the arcs, and a diagram of output has the parameter
valuesand possibly standard errors in parentheses
similarly associated with the arcs.) Convention 1 (Figure 1), due to McArdle (1980), most closely corresponds to the picture of a graph used in DAG theory.
It does not seem to have been widely adopted, possibly because it uses a nondirected arc between two
variables (latent or observable) to represent the covariance of their disturbance or error terms, and a
closed loop for variance. This could seem confusing
to some users, and logical to others. Convention 2
(Figure 2) graphs the disturbance and error terms,
together with the observed and latent variables, with
unit path coefficients from the former to the latter.
This is a complete counterpart of the equations of the
model. In this version, on the face of it, error terms are
exogenous latent variables. Convention 3 (Figure 3)
distinguishes disturbances and error terms from latent
variables. However, its mode of representation makes
it at least inconvenient to add values of error variances
or covariances (and their standard errors) to the diagram. Possibly this convention actively serves to discourage users from presenting error and disturbance
variances or from including disturbance covariances
in the path model they draw.
Of the 41 cases, 1 used Convention 1, 4 used Convention 2, 5 used Convention 3, and the rest exhibited
varying degrees of incompleteness, or in a few cases
did not follow any established convention. Of the 10
with complete path diagrams, 8 gave complete sets of
parameters. Of the 31 with incomplete path diagrams,
6 gave all parameters. A simple chi-square test gives
a significant association, but does not establish that
the use of an incomplete path diagram is the cause of
an incomplete account of the parameters!
As a recommendation, we simply repeat the truism
that the report should give all parameters and their
standard errors. Although the method of presentation
can be regarded as a matter of personal taste, it should
be consciously chosen with recognition of its advan-

REPORTING STRUCTURAL EQUATION ANALYSES

77

Figure 1. Path diagram, Convention 1: Disturbance variances modeled as closed,

nondirected arcs.

tages and disadvantages, and there should be no ambiguity from the readers viewpoint.

Alternative and Equivalent Models

For any set of multivariate data there will almost
always be more than one plausible structural model.
Both misfit and overfit of a single target model can be
taken to imply model modification (adding or deleting
parameters). Model modification indices, supplied by
a number of the available computer programs, can be
used to generate alternative models mechanically.
Seminal work by Stelzl (1986) and by Verma and
Pearl (1990) has shown how to generate alternative,
equivalent SEMs that cannot be distinguished empirically. MacCallum et al.s (1993) review should have
drawn the question of equivalent models to the attention of a wide range of SEM users, with a reasonably
immediate influence on the practice of SEM.
Tests of significance on differences in chi-squares
were used in 33 of the 41 studies to choose among
nested models; 6 of these were aided by Wald tests.
Only 4 of the 41 studies surveyed gave some recognition to the existence of equivalent models (2 from

1996 and 2 from 1997) from a set with 7 in 1995, 16

in 1996, and 17 in 1997. We thus add further evidence
to that given by MacCallum et al. (1993) of persistent
neglect of this question.
Endorsing remarks by Hoyle and Panter (1995) and
by Boomsma (2000), we note that desirably, plausible
competing models (nested or nonnested) should be
specified a priori by the investigator, along with the
preferred target model, and the relative goodness
of fit of those reported. Readers are reminded of an
important study by MacCallum, Roznowski, and
Necowitz (1992), which advised against taking a
quasi-random walk through a sequence of models
generated mechanically by model-modification indices. In view of earlier remarks advocating an examination of discrepancies as a basis for judging fit, we
note that this warning would also apply to the mechanical use of large standardized discrepancies for
this purpose. At the least, theoretical justification of
model modifications needs to be found and reported.
We suspect that post facto this task is commonly all
too easy. As MacCallum et al. pointed out, the prior
specification of a set of alternative models would generally be safer than model modification after the facts

78

MCDONALD AND HO

Figure 2. Path diagram, Convention 2: Disturbances modeled as latent variables.

(and, as noted, Hoyle & Panter, 1995, and Boomsma,

2000, are in agreement).
A more tolerant recommendation would allow the
researcher a few modifications of an initial model,
provided that a clear theoretical justification can be
found, and provided that a clear history of the decision steps is given to the reader, who may not share
the authors enthusiasm for their conclusions. Both
Hoyle and Panter (1995) and Boomsma (2000) specifically advised against adding disturbance covariance to improve fit. We endorse the general principle
that model modification should not be purely datadriven. However, as noted previously, the addition of
nondirected arcs does not change the causal model,
and a nondirected arc corresponds to a specified or
unspecified omitted variable. It is not unreasonable to
add such an arc if a plausible theory can be suggested
for it and further work is implied that explores the
possibility of measuring it.

Conclusion
We claim no special authority to offer a code of
practice for the reporting of SEM results. Recommendations, with varying degrees of confidence, have

been offered at the end of each section of this article.

Unquestionably, practice could be greatly improved
simply by a conscious intention on the part of investigators to make reasoned choices on each aspect of
their work and to report the basis of those choices.
Such choices include decisions to report sample covariances (or correlations with standard deviations)
and standardized discrepancies, or at the very least to
indicate means for the reader to access this information. The report should certainly include all the parameters and their standard errors. There should also
be a reasoned choice of a clear and complete form of
the path model structure, reported possibly as a conventional path diagram, and of the measurement
model, reported possibly in traditional tabular form.
Completeness is essential. The form of representation
remains a matter of taste, as long as it leaves no ambiguity. Again, a careful rational choice is needed for
conclusions about fit of the structural model, with
separate attention to its measurement and path components, and attention to individual discrepancies.
Global indices of fit, with criteria for their acceptability, do not in the present state of knowledge substitute
for a more detailed examination and careful judgment.
Investigators should also provide reasoned justifica-

REPORTING STRUCTURAL EQUATION ANALYSES

79

Figure 3. Path diagram, Convention 3: Disturbances distinguished from latent variables.

tions for omitted directed and nondirected arcs, which

jointly create a testable model, paying attention to
substantively plausible alternative and equivalent
models. An explicit account of the identifiability of
the model is very desirable, although in many cases
the careful reader will be able to determine this by
inspection.
SEM has been characterized as a dangerously conjectural technique for asking essential research questions which otherwise are impossible to consider
(McDonald, 1999, p. 367). This review indicates
some steps that can be taken to reduce the manifest
dangers accompanying its use.

References
Allison, S. (1987). Estimation of linear models with incomplete data. In C. C. Clogg (Ed.), Sociological methodology (pp. 71103). San Francisco: Jossey-Bass.
Amemiya, Y., & Anderson, T. W. (1990). Asymptotic chisquare tests for a large class of factor analysis models.
Annals of Statistics, 18, 14531463.
Anderson, J. C., & Gerbing, D. W. (1988). Structural equation modeling in practice: A review and recommended
two-step approach. Psychological Bulletin, 103, 411
423.

Anderson, J. C., & Gerbing, D. W. (1992). Assumptions

and comparative strengths of the two-step approach. Sociological Methods and Research, 20, 321333.
Anderson, T. W. (1989). Linear latent variable models and
covariance structures. Journal of Econometrics, 41, 91
119.
Arbuckle, J. L. (1997). Amos users guide. Chicago: SmallWaters.
Bekker, P. A., Merckens, A., & Wansbeek, T. J. (1994).
Identification, equivalent models, and computer algebra.
San Diego, CA: Academic Press.
Bentler, P. M. (1985). Theory and implementation of EQS:
A structural equations program. Los Angeles: BMDP
Statistical Software.
Bentler, P. M. (1990). Comparative fit indexes in structural
models. Psychological Bulletin, 107, 238246.
Bentler, P. M. (1995). EQS structural equations programs
manual. Encino, CA: Multivariate Software, Inc.
Bentler, P. M. (2000). Rites, wrong, and gold in model testing. Structural Equation Modeling, 7, 8291.
Bentler, P. M., & Bonett, D. G. (1980). Significance tests
and goodness-of-fit in the analysis of covariance structures. Psychological Bulletin, 88, 588606.
Bentler, P. M., & Dudgeon, P. (1996). Covariance structure
analysis: Statistical practice, theory, and directions. Annual Reviews of Psychology, 47, 563592.

80

MCDONALD AND HO

Bentler, P. M., & Weeks, D. G. (1980). Linear structural

equations with latent variables. Psychometrika, 45, 289
308.
Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley.
Bollen, K. A. (2000). Modeling strategies: In search of the
Holy Grail. Structural Equation Modeling, 7, 7481.
Bollen, K. A., & Arminger, G. (1991). Observational residuals in factor analysis and structural equation models.
In P. V. Marsden (Ed.), Sociological methodology (pp.
235262). Cambridge, MA: Blackwell.
Bollen, K. A., & Long, J. S. (Eds.). (1993). Testing structural equation models. Thousand Oaks, CA: Sage.
Boomsma, A. (2000). Reporting analyses of covariance
structures. Structural Equation Modeling, 7, 461483.
Breckler, S. J. (1990). Applications of covariance structure
modeling in psychology: Cause for concern? Psychological Bulletin, 107, 260273.
Browne, M. W. (1984). Asymptotically distribution free
methods for the analysis of covariance structures. British
Journal of Mathematical and Statistical Psychology, 37,
6283.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of
assessing model fit. In K. A. Bollen & J. S. Long (Eds.),
Testing structural equation models (pp. 136162). Thousand Oaks, CA: Sage.
Browne, M. W., & Du Toit, S. H. C. (1992). Automated
fitting of nonstandard models. Multivariate Behavioral
Research, 27, 269300.
Browne, M. W., Mels, G., & Cowan, M. (1994). Path
analysis: RAMONA: SYSTAT for DOS advanced applications (Version 6, pp. 167224). Evanston, IL: SYSTAT.
Browne, M. W., & Shapiro, A. (1988). Robustness of normal theory methods in the analysis of linear latent variate
models. British Journal of Mathematical and Statistical
Psychology, 41, 193208.
Chou, C. P., Bentler, P. M., & Satorra, A. (1991). Scaled
test statistics and robust standard errors for non-normal
data in covariance structure analysis: A Monte Carlo
study. British Journal of Mathematical and Statistical
Psychology, 44, 347357.
Curran, P. J., West, S. G., & Finch, J. F. (1996). The robustness of test statistics to nonnormality and specification error in confirmatory factor analysis. Psychological
Methods, 1, 1629.
Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977).
Maximum likelihood from incomplete data via the EM
algorithm. Journal of the Royal Statistical Society, Series
B, 39, 138.
Enders, C. K., & Bandalos, D. L. (2001). The relative per-

formance of full information maximum likelihood estimation for missing data in structural equation models.
Structural Equation Modeling, 8, 430457.
Fisher, F. M. (1970). A correspondence principle for simultaneous equation models. Econometrica, 38, 7392.
Fornell, C., & Yi, Y.-J. (1992). Assumptions of the two-step
approach to latent variable modeling. Sociological Methods and Research, 20, 291320.
Fraser, C., & McDonald, R. P. (1988). COSAN: Covariance
structure analysis. Multivariate Behavioral Research, 23,
263265.
Gollob, H. F., & Reichardt, C. S. (1987). Taking account of
time lags in causal models. Child Development, 58, 80
92.
Graham, J. W., & Hofer, S. M. (2000). Multiple imputation
in multivariate research. In T. Little, K. U. Schnabel, & J.
Baumert (Eds.), Modeling longitudinal and multilevel
data: Practical issues, applied approaches and specific
examples (pp. 201218). Mahwah, NJ: Erlbaum.
Hartmann, W. M. (1992). The CALIS procedure: Extended
users guide. Cary, NC: SAS Institute.
Hayduk, L. A., & Glaser, D. N. (2000a). Jiving the four-step
waltzing around factor analysis, and other serious fun.
Structural Equation Modeling, 7, 135.
Hayduk, L. A., & Glaser, D. N. (2000b). Doing the fourstep, right 2-3, wrong 2-3: A brief reply to Mulaik and
Millsap; Bollen; Bentler; and Herting and Costner. Structural Equation Modeling, 7, 111123.
Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrika, 45, 153161.
Heckman, J. J., & Robb, R. (1986). Alternative methods for
solving the problem of selection bias in evaluating the
impact of treatments on outcomes. In H. Wainer (Ed.),
Drawing inference from self-selected samples (pp. 63
107). New York: Springer.
Herting, J. R., & Costner, H. L. (2000). Another perspective
on the proper number of factors and the appropriate
number of steps. Structural Equation Modeling, 7, 92
110.
Hoyle, R. H., & Panter, A. T. (1995). Writing about structural equation models. In R. H. Hoyle (Ed.), Structural
equation modeling: Concepts, issues, and applications
(pp. 158176). Thousand Oaks, CA: Sage.
Hu, L., & Bentler, P. M. (1995). Evaluating model fit. In
R. H. Hoyle (Ed.), Structural equation modeling: Concepts, issues and applications (pp. 7699). Thousand
Oaks, CA: Sage.
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6, 155.

REPORTING STRUCTURAL EQUATION ANALYSES

Hu, L., Bentler, P. M., & Kano, Y. (1992). Can test
statistics in covariance structure analysis be trusted? Psychological Bulletin, 112, 351362.
Joreskog, K. G. (1973). A general method for estimating a
linear structural equation system. In A. S. Goldberger &
O. D. Duncan (Eds.), Structural equation models in the
social sciences (pp. 85112). New York: Seminar Press.
Joreskog, K. G., & Sorbom, D. (1989). LISREL 7: A guide
to the program and applications (2nd ed.). Chicago:
SPSS Inc.
Joreskog, K. G., & Sorbom, D. (1996). LISREL 8: Users
reference guide. Chicago: Scientific Software International.
Kano, Y. (1992). Robust statistics for test-of-independence
and related structural models. Statistics and Probability
Letters, 15, 2126.
Loehlin, J. C. (1992). Latent variable models: An introduction to factor, path, and structural analysis. Mahwah, NJ:
Erlbaum.
Lohmoller, J. B. (1981). LVPLS 1.6 program manual: Latent variables path analysis with partial least squares
estimation. Munich, Germany: Hochschule der Bundeswehr.
Long, J. S. (1983a). Confirmatory factor analysis: A preface to LISREL. Thousand Oaks, CA: Sage.
Long, J. S. (1983b). Covariance structure models: An introduction to LISREL. Thousand Oaks, CA: Sage.
MacCallum, R. C., & Austin, J. T. (2000). Applications of
structural equation modeling in psychological research.
Annual Review of Psychology, 51, 201226.
MacCallum, R. C., Roznowski, M., & Necowitz, L. B.
(1992). Model modifications in covariance structure
analysis: The problem of capitalization on chance. Psychological Bulletin, 111, 490504.
MacCallum, R. C., Wegener, D. T., Uchino, B. N., & Fabrigar, L. R. (1993). The problem of equivalent models in
applications of covariance structure models. Psychological Bulletin, 114, 185199.
Mardia, K. V. (1970). Measures of multivariate skewness
and kurtosis with applications. Biometrika, 57, 519530.
McArdle, J. J. (1980). Causal modeling applied to psychonomic systems simulation. Behavior Research Methods
and Instrumentation, 12, 193207.
McArdle, J. J. (1994). Structural factor analysis experiments
with incomplete data. Multivariate Behavioral Research,
29, 409454.
McArdle, J. J., & McDonald, R. P. (1984). Some algebraic
properties of the reticular action model for moment structures. British Journal of Mathematical and Statistical
Psychology, 37, 234251.
McDonald, R. P. (1978). A simple comprehensive model

81

for the analysis of covariance structures. British Journal

of Mathematical and Statistical Psychology, 31, 5972.
McDonald, R. P. (1989). An index of goodness-of-fit based
on noncentrality. Journal of Classification, 6, 97103.
McDonald, R. P. (1996). Path analysis with composite variables. Multivariate Behavioral Research, 31, 239270.
McDonald, R. P. (1997). Haldanes lungs: A case study in
path analysis. Multivariate Behavioral Research, 32,
138.
McDonald, R. P. (1999). Test theory: A unified treatment.
Mahwah, NJ: Erlbaum.
McDonald, R. P. (in press). What can we learn from the
path equations? Psychometrika.
McDonald, R. P., & Marsh, H. W. (1990). Choosing a multivariate model: Noncentrality and goodness of fit. Psychological Bulletin, 107, 247255.
McDonald, R. P., Parker, P. M., & Ishizuka, T. (1993). A
scale-invariant treatment for recursive models. Psychometrika, 58, 431443.
Meredith, W. (1964). Notes on factorial invariance. Psychometrika, 29, 177185.
Micceri, T. (1989). The unicorn, the normal curve and other
improbable creatures. Psychological Bulletin, 105, 156
165.
Mooijaart, A. (1993). Structural equation models with transformed variables. In K. Haagen, D. J. Bartholomew, &
M. Deistler (Eds.), Statistical modeling and latent variables (pp. 249258). Amsterdam: North Holland.
Mulaik, S. A., & Millsap, R. E. (2000). Doing the four-step
right. Structural Equation Modeling, 7, 3674.
Muthen, B. (1984). A general structural equation model
with dichotomous, ordered categorical and continuous latent variable indicators. Psychometrika, 49, 115132.
Muthen, B. (1989). Multiple group structural modeling with
non-normal continuous variables. British Journal of
Mathematical and Statistical Psychology, 42, 5562.
Muthen, B., & Kaplan, D. (1992). A comparison of some
methodologies for the factor analysis of non-normal
Likert variables: A note on the size of the model. British
Journal of Mathematical and Statistical Psychology, 45,
1930.
Muthen, B., Kaplan, D., & Hollis, M. (1987). On structural
equation modeling with data that are not missing completely at random. Psychometrika, 52, 431462.
Muthen, B., & Muthen, L. (1998). Mplus users guide. Los
Angeles, CA: Muthen & Muthen.
Neale, M. C. (1997). Mx: Statistical modeling (4th ed.).
Richmond, VA: Author.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems. San Mateo, CA: Morgan Kaufman.
Pearl, J. (1998). Graphs, causality, and structural equation

82

MCDONALD AND HO

models. Sociological Methods and Research, 27, 226

284.
Pearl, J. (2000). Causality: Models, reasoning and inference. Cambridge, England: Cambridge University Press.
Raykov, T., Tomer, A., & Nesselroade, J. R. (1991). Reporting structural equation modeling results in Psychology and Aging: Some proposed guidelines. Psychology
and Aging, 6, 499503.
Rubin, D. (1976). Inference and missing data. Biometrika,
63, 581592.
Rubin, D. (1987). Multiple imputation for nonresponse in
surveys. New York: Wiley.
Rubin, D. (1991). EM and beyond. Psychometrika, 56, 241
254.
Satorra, A., & Bentler, P. M. (1994). Corrections to standard errors in covariance structure analysis. In A. von
Eye & C. C. Clogg (Eds.), Latent variable analysis: Applications to developmental research (pp. 399419).
Thousand Oaks, CA: Sage.
Scheines, R., Spirtes, P., Glymour, C., Meek, C., & Richardson, T. (1998). The TETRAD project: Constraintbased aids to causal model specification. Multivariate
Behavioral Research, 33, 65117.
Shapiro, A., & Browne, M. W. (1987). Analysis of covariance structures under elliptical distributions. Journal of
the American Statistical Association, 82, 10921097.
Spirtes, P., Richardson, T., Meek, C., Scheines, R., & Glymour, C. (1998). Path diagrams as a structural equation
modeling tool. Sociological Methods and Research, 27,
182225.
Steiger, J. H. (1988). Aspects of personmachine communication in structural modeling of correlations and covariances. Multivariate Behavioral Research, 23, 281290.
Steiger, J. H. (1995). SEPATH: STATISTICA version 5.
Tulsa: StatSoft, Inc.

Steiger, J. H., Shapiro, A., & Browne, M. W. (1985). On the

multivariate asymptotic distribution of sequential chisquare statistics. Psychometrika, 50, 253264.
Stelzl, I. (1986). Changing a causal hypothesis without
changing the fit: Some rules for generating equivalent
path models. Multivariate Behavioral Research, 21, 309
331.
Tucker, L. R., & Lewis, C. (1973). A reliability coefficient
for maximum likelihood factor analysis. Psychometrika,
38, 110.
Verma, T., & Pearl, J. (1990). Equivalence and synthesis of
causal models. In Proceedings of the fifth conference on
uncertainty in artificial intelligence (pp. 220227). Amsterdam: Elsevier.
West, S. G., Finch, J. F., & Curran, P. J. (1995). Structural
equation models with non-normal variables. In R. H.
Hoyle (Ed.), Structural equation modeling: Concepts, issues and applications (pp. 5675). Thousand Oaks, CA:
Sage.
Wiley, D. E. (1973). The identification problem for structural equation models with unmeasured variables. In
A. S. Goldberger & O. D. Duncan (Eds.), Structural
equation models in the social sciences (pp. 6983). New
York: Seminar Press.
Wright, S. (1921). Correlation and causation. Journal of
Agricultural Research, 20, 557585.
Yuan, K., Chan, W., & Bentler, P. F. (2000). Robust transformation with applications to structural equation modeling. British Journal of Mathematical and Statistical
Psychology, 53, 3150.

Received May 12, 2000

Revision received October 29, 2001
Accepted November 1, 2001